Adaptive signal control and coordination for urban traffic control in a connected vehicle environment: A review

Jiangchen Li; Liqun Peng; Kaizhe Hou; Yong Tian; Yulin Ma; Shucai Xu; Tony Z. Qiu; Jiangchen Li; Liqun Peng; Kaizhe Hou; Yong Tian; Yulin Ma; Shucai Xu; Tony Z. Qiu

doi:10.48130/DTS-2023-0008

Signal control	Data type	Traffic prediction	Control strategy
Fixed-time	Historical	N/A	Pre-defined timing plans
Actuated	Real-time	N/A	Simple logics
Adaptive	Real-time	Predictions by traffic models	Signal optimizations

Table 1.

Summary of three conventional signal control systems.

Category	Adjusted control	Responsive control	Advanced adaptive control
^a Data quality: sensor density level (L)	Static sensor data
^a Data quality: sensor density level (L)	L1 & L1.5, less than one sensor up to one sensor per link	L2, one sensor per link up to one per lane	L3, two sensors per lane
^a Responsive to demand variations	Slow reactive response based on pre-calculated historical traffic flow	Prompt reactive response based on changes in regularly disrupted traffic	Very rapid proactive response based on short-term predicted movements
^a Change frequency in control plan (HZ)	Minimum of 15 minutes, usually several times during a rush period, (< 1/900 HZ)	Minimum of 5-15 minutes, per several cycles, (< 1/300 HZ)	Continuous adjustments are made to all timing parameters, per several seconds (< 1/5 HZ)
^c Control strategy	Pattern matching from pre-stored plans by static optimization	Cyclic timing plan generating and matching via static/dynamic optimization	Real-time timing adjusting via dynamic optimization and optimal control
^a,b Generations of UTCSs (G)	G1 & G1.5^a , e.g., SCATS^[28]	G2^a, e.g., SCOOT^[27]	G3^b , e.g., OPAC^[29], RHODES^[30], ACS Lite^[53]
Coordination included	Mostly yes	Mostly yes	Yes
^a Adopted from Klein et al.^[14] and Stevanovic et al.^[15]. ^b Summarized from Gartner et al.^[52] and Wang et al.^[16]. ^c Identified in this report drawn from across a number of studies.

Table 2.

Fine classifications of adaptive signal control (ASC)^[14–16,52].

Category	Adjusted control	Responsive control		Advanced adaptive control
^a Data quality: sensor density level (L)	Same as Table 2
^a Responsive to demand variations
^a Change frequency in control plan
^c Control strategy
^a,b Generations of UTCSs (G)
Specific control strategy for Coordination	Advancement of the quality of progression, e.g., classical MAXBAND^[33] and recent AMBAND^[68]		Optimization of a performance index, e.g., MITROP^[34]

Table 3.

Classifications of signal coordinations in UTCSs^[19].

Author, year	Objective functions⁺
Author, year	Delay¹	Queue length²	Waiting time³	Stop⁴	Travel time⁵	Type
Gradinescu et al.^[75] in 2007	Average delay					1
Chou et al.^[76] in 2012	Vehicle and Passenger delay			Stops		*
Nafi and Khan^[77] in 2012			Average waiting time			3
Chang and Park^[78] in 2013		Queue length	Junction waiting time			*
Ahmane et al.^[79] in 2013		Queue length				2
Cai et al.^[80] in 2013					Travel time	5
Pandit et al.^[81] in 2013	Delay					1
Lee et al.^[82] in 2013					Cumulative Travel time	5
Kari et al.^[83] in 2014	Travel delay					1
Guler et al.^[36] in 2014	Total delay			Stops		*
Tiaprasert et al.^[84] in 2015		Queue length				2
Feng et al.^[1] in 2015	Vehicle delay	Queue length				1
Younes and Boukerche^[85] in 2016	Delay					1
Feng et al.^[32] in 2016	Vehicle delay					1
Islam et al.^[88] in 2017		Queue length				2
Liu et al.^[39] in 2017			Average waiting time			3
Cheng et al.^[86] in 2017			Average waiting time			3
Feng et al.^[87] in 2018	Total delay					1
Ban et al.^[89] in 2018	Delay					1
Al Islam et al.^[90] in 2020	Average delay				Total travel time	*
Li et al.^{[91, 92]} in 2021	Delay					1
Mo et al.^[93] in 2022	Average delay					1
⁺ Index type: 1 delay, 2 queue length, 3 waiting time, 4 stop, 5 travel time, * combination.

Table 4.

Summary of the objective functions in the existing CV-based ASCs applied to both the isolated intersection and multiple intersections.

Category	CV-based basic ASC	CV-based advanced ASC
^{a, c} Data quality: sensor density level (L) and market penetration rate (P_cv)	Mobile sensor data
	L4^a, P_cv = 100%, i.e., 100 % market penetration rate	L3.5^c & L4^a, P_cv < 100% & P_cv = 100%, i.e., both non-full and full market penetration rate
	Each connected vehicle (CV) regularly reports its location, speed, and possibly its destination^a
^b Responsive to demand variations	Very rapid proactive response based on short-term traffic predictions
^b Change frequency in control plan (HZ)	Continuous adjustments, per several seconds to per second (< 1 HZ)
^c Control Strategy	Real-time timing adjustment via static optimization, dynamic optimization, and optimal control
^c Generations of UTCSs (G)	G4^c, e.g., work by Gradinescu et al.^[75]	G4.5^c , e.g., PAMSCOD^[42] and detector-free ASC^{[ 87]}
^a Adopted from Klein et al.^[14], Stevanovic et al. ^[15]. ^b Summarized from Gartner et al. ^[52], and Wang et al. ^[16]. ^c Identified in this report.

Table 5.

Fine classifications of the CV-based ASC^[14−16,52].

Author, year	Country/ region	Institution
He et al.^[42] in 2012	USA	University of Arizona
C.M. Day et al.^[40] in 2016	USA	Purdue University
Li et al.^[41] in 2016	USA	Purdue University
Feng et al.^[32] in 2016	USA	University of Arizona
Beak et al.^[19] in 2017	USA	University of Arizona, University of Michigan
C.M. Day et al.^[98] in 2017	USA	Purdue University
Remias et al.^[46] in 2018	USA	Purdue University
Zheng et al.^[99] in 2018	USA, China	University of Michigan, Didi Chuxing LLC
Mo et al.^[93] in 2022	USA	Columbia University

Table 6.

Summary of the CV-based advanced signal coordination systems’ research teams and outputs.

Category	CV-based advanced signal coordination systems
^{a, c} Data quality: sensor density level (L) and market penetration rate (P_cv)	Mobile sensor data
	L3.5^c & L4^a, P_cv < 100% & P_cv = 100%, i.e., both non-full and full market penetration rate
^b Responsive to demand variations	Slow reactive response based on historic traffic flows	Rapid proactive response based on short-term predicted movements
^b Change frequency in control plan (HZ)	Minimum of 15 min−3 h, (< 1/900 HZ)	Continuous adjustments, usually per cycle (< 1/100 HZ)
Minimum P_{cv_min}	0.1% for per 3 hrs change, 5% for per 15 mins change	25% for per cycle change
Specific control strategy of coordination	offline offset method, e.g., detector-free method ^{[98,40,41,46]}	online priority-based method, e.g., adaptive coordination method^[19,32,42]
^c Generations of UTCSs (G)	UTCS G4.5^c
^a Adopted from Klein et al.^[14], Stevanovic et al. ^[15]. ^b Summarized from Gartner et al.^[52], and Wang et al.^[16]. ^c Identified in this report.

Table 7.

Fine classifications of the CV-based advanced signal coordination systems^{[19,32,40−43,98]}.

Category	Non-CV-based Adjusted control	Non-CV-based Responsive control	Non-CV-based Advanced adaptive control	CV-based Basic ASC	CV-based Advanced ASC
^a Data quality: sensor density level (L)	Static sensor data			Mobile sensor data
^a Data quality: sensor density level (L)	L1 & L1.5, less than one sensor up to one sensor per link	L2, one sensor per link up to one per lane	L3, two sensors per lane	L4^a, P_cv = 100%, i.e., 100% market penetration rate	L3.5^c & L4^a, P_cv < 100% & P_cv = 100%, i.e., both non-full and full market penetration rate
^a Responsive to demand variations	Slow reactive response based on pre-calculated historical traffic flow	Prompt reactive response based on changes in regularly disrupted traffic	Very rapid proactive response based on short-term predicted movements	Very rapid proactive response based on short-term traffic predictions
^a Change frequency in control plan (HZ)	Minimum of 15 min, usually several times during a rush period, (< 1/900 HZ)	Minimum of 5−15 min, per several cycles, (< 1/300 HZ)	continuous adjustments are made to all timing parameters, per several seconds (< 1/5 HZ)	Continuous adjustments, per several seconds to per second (< 1 HZ)
^c Control strategy	Pattern matching from pre-stored plans by static optimization	Cyclic timing plan generating and matching via static/dynamic optimization	real-time timing adjustment via dynamic optimization and optimal control	Real-time timing adjustment via static optimization, dynamic optimization, and optimal control
^a,b Generations of UTCSs (G)	G1 & G1.5^a, e.g., SCATS^[28]	G2^a, e.g., SCOOT^[27]	G3^b, e.g., OPAC^[29], RHODES^{[ 30]}, ACS Lite^[53]	G4^c, e.g., the work by Gradinescu et al.^[75]	G4.5^c, e.g., PAMSCOD^[42] and detector-free ASC^[87]
Coordination included	Mostly yes	Mostly yes	Yes	Mostly yes	Mostly yes
Traffic model	Microscopic/ macroscopic/ mesoscopic models			Mostly microscopic models
^* Summarized from previous Tables 2 & 5, where further details of the above notations are available.

Table 8.

Fine classifications of traditional (non-CV-based) and CV-based ASC*.

Category	Non-CV-based Adjusted control	Non-CV-based Responsive control		Non-CV-based Advanced adaptive control	CV-based Advanced signal coordination systems
^a Data quality: sensor density level (L)	Static sensor data				Mobile sensor data
^a Data quality: sensor density level (L)	L1 & L1.5,	L2,		L3,	L3.5^c & L4^a, P_cv < 100% & P_cv = 100%, i.e., both non-full and full market penetration rate
^a Responsive to demand variations	Same as Table 8				Slow reactive response based on historical traffic flows	Rapid, proactive response based on short-term predicted movements
^a Change frequency in control plan (HZ)					Minimum of 15 min−3h, (< 1/900 HZ)	Continuous adjustments, usually per cycle (< 1/100 HZ)
^c Minimum P_{cv_min}					0.1% for per 3 hrs change, 5% for per 15 mins change	25% for per cycle change
^a,b Generations of UTCSs (G)					G4.5^c ,

Specific control strategy for Coordination	Advancement of quality of progression, e.g., classical MAXBAND^[33] and recent AMBAND^[68]		Optimization of a performance index, e.g., MITROP^[34]		Offline offset method, e.g., detector-free method^{[98,40,41,46]}	Online priority-based method, e.g., adaptive coordination method^[19,32,42]
Traffic model	Microscopic/ macroscopic/ mesoscopic models				Mostly microscopic models
^* Summarized from previous Tables 3 & 7, where further details of the above notations are available.

Table 9.

Fine classifications of traditional (non-CV-based) and CV-based signal coordination*.

Data Type		Spatial-temporal property of traffic data	Cost^*	Extra proactive data	Pros/ Cons
Static sensor data		Instantaneous data at fixed location	High	No	Cons
Mobile sensor (CV) data	Full penetration	Complete spatial and temporal CV data, high frequency of data exchange	Low	Yes, e.g., priority request data	Pros
	Low penetration	Limited CV data			Cons
	Low penetration	Large missing of non-CV data			Cons
^* Usually considering the installation and maintenance cost.

Table 10.

Summary of the data comparisons and limitations for both the static and mobile sensor data.

Low penetration rate issue	Limited CV data issue	Missing of non-CV data issue	CV applications	Min P_cv
Low penetration rate issue	Proposed methods		CV applications	Min P_cv
Goodall et al.^[94] in 2014	n/a	Micro-simulation-based estimation of the non-CV position	CV-ASC	10%−25%
Feng et al.^[1] in 2015	n/a	EVLS algorithm	CV-ASC	25%−50%
Day et al.^{[98, 40, 41, 46]} from 2016 to 2018	Historical limited CV data-based aggregation	n/a	detector-free coordination	5%, 15 mins change
Beak et al.^[19] in 2017	Stop-bar detector-assisted method	n/a	adaptive coordination	25%
Feng et al.^[87] in 2018	Probabilistic model based on both prior arrival distribution and historical CV data	n/a	CV-ASC	10%
Al Islam et al.^{[90 ]} in 2020	Spatial vehicle distribution estimation by CVs	vehicle trajectories via both the loop and CV data	CV-ASC and coordination	0%, 10%
Li et al.^{[91, 92]} in 2021	Vehicle-triggered platoon dispersion	n/a	CV-based coordination	5%, 10%
Mo et al.^[93] in 2022	Decentralized learning method	n/a	CV-ASC	10%
Zhang et al.^[104] in 2022	Bayesian deduction	n/a	CV-ASC	5%, 10%

Table 11.

Summary of studies targeting the low-penetration issue for urban signals.

Decade	Typical UTCSs	Data^a	Global optimization formulation and/or solution algorithm	Traffic model
1960s	TRANSYT in UK in 1968	Loop data	Domain-constrained optimization	DSM model^[15]
1970s	SCATS in Australia in 1979	SL, Loop data	Strategic and tactical control	Flow-delay profiles^[15]
	SCOOT in UK in 1979	US, Loop data	Domain-constrained optimization	Flow-occupancy profiles, DSM model^[15]
	DYPIC in UK in 1974 ^[108]	US, Loop data	Backward dynamic programming^[108], Rolling horizon approach	DSM model
1980s -1990s	OPAC in US in 1983^[15]	MB & SL, Loop data	Complete enumeration / exhaustive enumeration^{[111, 112]}, Rolling horizon approach	DSM model^[15]
	RHODES in US in 1992^{[ 15]}	MB & SL, Loop data	Dynamic programming^{[111, 112]}, Rolling horizon approach^[30]	DSM model^[15]
	UTOPIA /SPOT in Italy in 1985^[15]	US & SL, Loop data	Online dynamic optimization and off-line optimization^[108] , Rolling horizon approach^[113]	DSM model
	PRODYN in France in 1984^[108]	US, Loop data	Forward dynamic programming^{[111, 112]} , Rolling horizon approach^[109]	DSM model
2000s	ACS-lite in US in 2003^[15]	US, Loop data	Domain-constrained optimization, three levels of optimization methodology	DSM model
2010s	Aboudolas et al. in 2010^[109]	AL, Loop data	Quadratic programming, Rolling horizon approach	SFM model
	Liu & Qiu in 2016^[110]	US & SL, Loop data	Multi-objective optimization problem and an evolutionary algorithm	Extended SFM model
	Hao et al. in 2018^{[114, 115]}	US, Loop data	Model predictive control-based method integrating optimizations	CTM model
	Han et al. in 2018^[116]	n/a	Linear quadratic model predictive control	Extended CTM model
	Lu et al. in 2019^[117]	n/a	Explicit model predictive control	SFM model
	Pedroso and Batista in 2021^[118]	US	Decentralized and decentralized-decoupled traffic-responsive urban control	Decentralized SFM
	Souza et al. in 2022^[119]	Loop data, Historical data	Integrating signal control and routing	Multi-commodity SFM
^a SL = stop-line, MB = mid-block, US = upstream, AL = arbitrary location, adopted from Stevanovic et al. ^[15] and Aboudolas et al.^[109].

Table 12.

Summary of traditional UTCSs applied different traffic models.

Typical UTCSs	Data^a	Rolling horizon approach	Global optimization formulation and/or solution algorithm
OPAC^[15]	MB & SL, Loop data	Yes^[15]	Complete enumeration (CE) / exhaustive enumeration^{[111, 112]}
RHODES^[15]	MB & SL, Loop data	Yes^[30]	Dynamic programming^{[ 111, 112]}
UTOPIA/SPOT^[15]	US & SL, Loop data	Yes^[113]	Online dynamic optimization and off-line optimization^{[ 108]}
PRODYN^[108]	US, Loop data	Yes^[109]	Forward dynamic programming^{[111, 112]}
DYPIC^{[ 108]}	US, Loop data	Yes^{[ 108]}	Backward dynamic programming^{[ 108]}
Aboudolas et al.^[109] in 2010	AL, Loop data	Yes	Quadratic programming
Liu & Qiu^[110] in 2016	US & SL, Loop data	Yes	Multi-objective optimization problem and an evolutionary algorithm
Hao et al.^{[114, 115]} in 2018	US, Loop data	Yes	MPC-based method integrating optimizations, CTM model
Jamshidnejad et al.^[141] in 2018	Loop data	Yes	Sustainable model-predictive control, S-model
Han et al.^[116] in 2018	Loop data	Yes	LQ-MPC, extended CTM, corridor
Lu et al.^[117] in 2019	Loop data	Yes	Explicit model predictive control (EMPC), SFM model
Pedroso & Batista^[118] in 2021	Loop data	One-step	Decentralized and decentralized-decoupled traffic-responsive urban control, Decentralized SFM
Souza et al.^[119] in 2022	Loop data	Yes	Integrating signal control and routing, Multi-commodity SFM
^a SL = stop-line, MB = mid-block, US = upstream, AL = arbitrary location, adopted from Stevanovic et al. ^[15] and Aboudolas et al. ^[109].

Table 13.

Summary of traditional UTCSs using the rolling horizon approach.

Typical works	Platform^a	Intelligent strategy	Control features
Jin et al.^[50] in 2017	Embedded device	Fuzzy-based	A fuzzy group-based approach, machine-to-machine connectivity
Gao et al.^[51] in 2017	Centralized structure	Deep reinforcement learning-based	Convolutional neural network for automatic feature crafting, experience replay and target network for stability
Wu et al.^[143] in 2019	Edge computing	Deep reinforcement learning-based	Distributed reinforcement learning
Zhou et al.^[144] in 2021	Hierarchical structure	Deep reinforcement learning-based	Multi-agent training with rich CV data, hierarchical control
Zhang et al.^[145] in 2021	Edge computing	Deep reinforcement learning-based	Multi-agent actor-critic control, value decomposition, a cooperative scheme
Wu et al.^[146] in 2022,	Edge computing	Deep reinforcement learning-based	Game theory-aided reinforcement learning
Wang et al.^[147] in 2022	Edge computing	Deep reinforcement learning-based	Social features and connections
Mo et al.^[93] in 2022	Decentralized	Deep reinforcement learning-based	Asymmetric advantage actor-critic, non-CV, and CV data for offline training, CV data for online control
Chen et al.^[142] in 2022	Edge, distributed	Distributed dynamic route-based	Distributed backpressure principle, dynamic route control

Table 14.

Summary of UTCSs using modern intelligent approaches.

Authors	CV data	Rolling horizon approach	Global optimization formulation and/or solution algorithm*	CV applications	Benefit⁺
Gradinescu et al.^[75] in 2007	Online	No	Static optimization¹	CV-ASC	28.3%¹
Priemer et al.^[161] in 2009		No	Dynamic optimization with DP & Complete enumeration^2a	CV-ASC	24%¹
Lee et al.^[82] in 2013		No	Static optimization¹	CV-ASC	34%⁵
Cai et al.^[80] in 2013		No	Dynamic optimization^2c	CV-ASC	11.69%⁵
Pandit et al.^[81] in 2013		No	Dynamic optimization^2c	CV-ASC	~25%¹
Kari et al.^[83] in 2014		No	Static optimization¹	CV-ASC	57.31%¹
Guler et al.^[36] in 2014		No	Dynamic optimization^2c	CV-ASC	~50%^*
Younes et al.^[85] in 2016		No	Scheduling algorithm^2c	CV-ASC	25%¹
Islam et al.^[88] in 2017		No	Modified MILP¹	CV-ASC	27%²
Liu et al.^[39] in 2017		No	Reinforcement learning^2c	CV-ASC	~30%³
PAMSCOD^[42] and its variant ^[45] in 2012 and 2014, respectively		Yes	MILP^2b	CV-ASC	25%¹
Goodall et al.^[35] in 2013		Yes^[16]	Dynamic optimization with rolling horizon^2b	CV-ASC	20%⁴
Feng et al.^[1] and its variant^[87] in 2015 and 2018, respectively		Yes	Hybrid structure^2b	CV-ASC	16.33%¹
C.M. Day et al.^{[98, 40, 41, 46]} from 2016 to 2018	Offline	No	Static optimization¹	CV-based coordination	−
Priority-based method ^[32]^[42] in 2016	Online	No	Static optimization¹	CV-based coordination	25%¹
Beak et al.^[19] in 2017	Online	No	Static optimization¹	CV-based coordination	19%¹
Al Islam et al.^[90] in 2020	Online	Yes	Dynamic optimization with rolling horizon^2b	CV-ASC and coordination	50%^*
Li et al.^{[91, 92]} in 2021	Online	Yes	MPC³	CV-based coordination	35%¹
Zhang et al.^[104] in 2022	Offline & online	No	Deep reinforcement learning-based^2c	CV-ASC	66%¹
Mo et al.^[93] in 2022	Offline & online	No	Deep reinforcement learning-based^2c	CV-ASC	30%¹
^* 1 = static optimization-based control, 2a = dynamic optimization-based control with the DP, 2b = dynamic optimization-based control with the rolling horizon scheme, 2c = dynamic optimization-based control with other intelligent approaches, 3 = MPC. ⁺ index type: 1 delay, 2 queue length, 3 waiting time, 4 stop, 5 travel time, * combination.

Table 15.

Summary of CV-based signal control systems.